🧩 Why the interview matters more than any technique
New here? Here “context" isn’t just any file — it’s the set of information the agent reads before acting: who you are, what your business does, who else is involved in decisions, what’s urgent today. Without this, the agent has no basis for making a judgment — it just gives generic answers, the kind any chatbot gives to anyone.
In Phase 1 (module 7.1), you set up the house: the folder, the CLAUDE.md
root, the directory structure. That’s the skeleton—necessary, but empty. Phase 2 is what fills this house with people: who lives there, what each person does, and what matters to each one. It’s the difference between giving a new employee an empty office and sitting with them for half an hour to explain how the company really works.
💡 Main Concept
Technique without context is a polite but useless agent. Context without technique is a notebook. Together, they’re what separates “answering questions” from “taking action for you.”
- •An agent without context answers “it depends on your situation” to everything—because it doesn’t know your situation.
- •An agent with real context already knows you prefer short meetings, that your business partner approves purchases over R$ 2 mil, and that you don’t send urgent messages on Friday afternoons.
🙋 Area 1 — Who you are
The first area of the interview isn’t about the business — it’s about you: your role, how you work, what you value. An assistant that doesn’t know whether you’re the owner making every decision alone or a manager who reports to someone else will get the tone wrong in every recommendation it makes.
🔍 Example questions
- Role: "Do you make the final decisions on day-to-day matters, or is there someone above you who needs to approve them?"
- Work style: "Do you prefer to review each step before taking action, or see the finished result and make corrections afterward?"
- Values: "Is there anything you would NEVER accept being done in your name, even if it technically worked? (e.g., promising a deadline that can’t be met, bad-mouthing a competitor)"
- Real routine: "What time of day do you usually make important decisions, and what time do you just want to be left alone?"
💡 Practical Tip
Vague answers like “I like things done well” don’t help at all. Ask for a concrete example: “Tell me about a time someone did a task the wrong way—what exactly was wrong?” The concrete answer is material the assistant can actually use.
🏢 Areas 2 and 3 — The business and the team
After you, the assistant needs to understand what makes the business work: what it sells, who it sells to, and how the money comes in. This isn't a company profile—it's a practical explanation, the way you'd explain it to a 15-year-old nephew who's never heard of your industry.
And right after that comes the team: who else is involved in decisions, and especially—who decides what. This prevents a classic mistake: the assistant “approving” something on your behalf when it was actually someone else's decision.
Caption: each question answered for real becomes a file inside contexto/, and it’s this file—not the conversation that has already passed—that the agent rereads before each new task.
🔍 Example questions
- The deal: "If I were to explain your business to someone in one sentence, what would have to be included? And how does the money come in — direct sales, subscription, commission?"
- For whom: "Who is your typical customer — can you describe in two or three sentences who usually buys from you?"
- The team: "Besides you, who else is involved day to day? What does each person handle?"
- Who decides what: "Is there any decision only you make and never delegate? And any decision someone else makes, even if you stay involved?"
⏱️ Areas 4 and 5 — Priorities and communication
New here? Urgent e important aren't the same thing: urgent means something that needs a response today (an angry customer, an account coming due); important means something that changes the direction of the business even without urgency (planning the next quarter). An assistant that confuses the two becomes a false alarm siren — or, worse, ignores what really deserved attention.
The communication area defines the assistant’s “voice”: when it should summarize in one line and when it should provide details; whether the tone is formal or informal; whether it sends a message at the end of the day or only when you ask.
🔍 Example questions
- Priorities: "What's on fire this week? And what's important but can wait a month?"
- Real urgency: "What in your routine really deserves to interrupt you right away — and what can wait until the end of the day?"
- Tone: "When you read a summary, do you prefer it straight to the point ('3 days left, take care of it') or do you prefer all the context before the conclusion?"
- When to be brief vs. detailed: "For what kinds of topics do you want just the result, and for what kinds do you want to see the full reasoning before you trust it?"
✓ Helpful response
- ✓"The only truly urgent things are a customer complaint or a payment due today."
- ✓"Send me a 3-line summary at 6 p.m.; if it's a money decision, contact me right away."
✗ Unhelpful response
- ✗"Everything is kind of urgent, it depends."
- ✗"I prefer that it use common sense" (common sense doesn't become a file — it becomes guesswork).
🔐 Area 6 — What to delegate, and the prompt that guides the interview
The last area is the most delicate: what the assistant is allowed to do alone, and what needs your approval before it starts executing. This connects directly to the permissions you configured in module 0.6—here you translate those technical rules into business language: "it can reply to simple emails on its own, but budgets over R$ 500 only after I approve."
🔍 Example questions
- "What kind of task would you trust completely, without reviewing it first?"
- "What kind of task must NEVER go out without your approval — even if it seems obvious?"
- "Is there a dollar amount that serves as a dividing line (above this amount, always get approval)?"
Now, the practical part of this phase: a complete prompt, ready to paste into Claude Code, so it can conduct this interview with you—question by question, area by area—and save the context files in the right place.
Você vai conduzir comigo uma entrevista estruturada para gerar os arquivos de
contexto do meu assistente digital, dentro da pasta contexto/ do meu projeto
.
Regras da conversa:
- Faça UMA pergunta de cada vez. Espere minha resposta antes de seguir.
- Se minha resposta ficar vaga ou genérica, peça um exemplo concreto antes
de continuar (ex.: "me dá um caso real que isso aconteceu").
- Não invente nada que eu não te disse. Se eu não souber responder algo
agora, anote como "em aberto" e siga em frente.
- Ao final de cada área, mostre um resumo do que você entendeu e pergunte
se está correto antes de salvar.
Cubra estas 6 áreas, nesta ordem, uma pergunta por vez dentro de cada:
1. QUEM SOU EU — meu papel, meu jeito de trabalhar, meus valores e o que eu
nunca aceitaria que fosse feito em meu nome.
2. O NEGÓCIO — o que ele faz, para quem, como o dinheiro entra.
3. O TIME — quem mais participa, quem decide o quê.
4. PRIORIDADES — o que é urgente essa semana vs. o que é importante mas
pode esperar.
5. COMUNICAÇÃO — tom, formalidade, quando eu quero resumo curto e quando
quero o detalhe todo.
6. O QUE DELEGAR — o que você pode fazer sozinho, e o que precisa da minha
aprovação antes de agir (inclua uma linha de valor em dinheiro se eu
te der uma).
Ao terminar cada área, salve um arquivo Markdown separado dentro de
contexto/ (ex.: contexto/quem-sou-eu.md, contexto/negocio.md,
contexto/time-decisoes.md, contexto/prioridades.md, contexto/comunicacao.md,
contexto/delegacao.md). NÃO junte tudo em um arquivo único — arquivos
separados por área facilitam eu revisar e atualizar cada um depois.
Antes de escrever qualquer arquivo, me avise que vai escrever e mostre
o conteúdo. NUNCA inclua senha, chave de API, token ou dado pessoal de
terceiros nesses arquivos — se eu mencionar algo assim sem querer, avise
e sugira remover.
Comece pela área 1, pergunta 1. How to verify: at the end, open the folder contexto/ and check that there are 6 separate files (one per area), each with concrete — not generic — answers and no sensitive data.
🔒 Privacy notice — required
These files live in a local folder on your computer — they do NOT get uploaded anywhere unless you decide to (for example, if you make a cloud backup yourself in module 7.5). Even so, never put these in them:
- ✗Passwords, API keys, access tokens — put these in an environment variable (module 4.4), never in plain text.
- ✗Third parties’ personal data without consent (CPF, address, customers’ or employees’ health information) — even if it seems like "just an example".
- ✗Anything you wouldn't show a new employee on their first day.
🔁 The interview never really “ends”
The first round is the longest — it covers all 6 areas at once, at a comfortable pace. But your business changes: new people join the team, one quarter’s priorities aren’t the next quarter’s, and you realize you delegated too much (or too little) of a task. The interview returns in focused reviews, not repeated marathons.
First round — broad
All 6 areas completed, using this module’s prompt. Set aside about 40 minutes and take your time—this is where most of the value comes from.
First few weeks — fine-tuning
In practice, you realize the daily summary is too long, or that the assistant asked for approval for something it could have done on its own. Go back to the file for that specific area and adjust it.
Business milestones — occasional review
Did you hire someone, switch providers, or enter a peak season? Time to review time-decisoes.md or prioridades.md — not the entire file, just the area that changed.
Continuous—never “closed”
There is no final version. The context evolves with the business—that’s exactly what Phase 5 (Backup and Versioning, Module 7.5) helps keep organized over time.
✓ Well-conducted interview
- ✓Specific answers, with a real example (“customer X called to complain about a delay”)
- ✓6 separate files organized by area — easy to find and update
- ✓Clear statement of what needs approval
✗ Rushed interview
- ✗Generic answers (“I like everything organized”) without an example
- ✗A single giant file mixing all 6 areas
- ✗No approval rule — by default, the assistant decides everything on its own
💡 Practical Tip
After the first round, reread the 6 files out loud as if you were explaining them to a real employee. If any sentence sounds strange when spoken aloud, it’s probably too vague for the agent too.
Quick check (optional): Why separate context files by area instead of using a single file?